Every provider has a dated governance record separating automated retrieval, attributed factual comparison, content reproduction, bulk redistribution and paid resale.
Methodology
How the calculators are built
Every page is based on published source material, explicit exclusions, and review dates that make it easier to judge what the estimate does and does not cover.
Source-use gate
A working parser is not permission to publish
If published provider terms block automated access, Guardian stops before any network request. Existing observations remain audit-only and cannot enter recommendations.
The source-use review expires on a short schedule. Reactivation after a terms block requires written permission or qualified legal review with dated evidence.
Workload benchmark layer
How published performance becomes cost
Only exact MLPerf 6.0 results from the official MLCommons repositories enter the benchmark catalog. Every snapshot records a full commit SHA, source paths and content hashes.
Training time uses ten successful runs per profile. The fastest and slowest are discarded, then the remaining eight are averaged, matching the published methodology.
Performance evidence and pricing evidence stay separate. The current price is joined only when provider, GPU family and market match; the benchmark result is never generalized to another workload.
Rules
Base rules for every calculator
Do not publish numbers that cannot be checked against a source page.
Prefer official pricing pages and explicit provider documentation over inferred values.
Mark missing or uncertain data rather than pretending it is exact.
Common exclusions
What most calculations leave out on purpose
Storage, egress, taxes, and reserved discounts are excluded unless a provider publishes them clearly and comparably.
Per-GPU equivalents are labelled when a provider only publishes pricing for a multi-GPU node.
Only pages with live source links and visible review dates are kept in the public site.
Automation
How updates are kept from drifting
Authorized official GPU pricing snapshots are refreshed by the Guardian pipeline before public data is committed.
Contract validation, pricing guardrails, Astro build, and SEO smoke checks must pass before automated data updates are pushed.
If source rights fail, a fetch fails, or a price falls outside expected bounds, the affected records are suspended or quarantined instead of silently becoming current data.